MiniMax Unveils M1 First Open-Source Hybrid AI Model for Complex Tasks

Beijing, China – MiniMax, a leading Chinese artificial intelligence company, has announced the release of MiniMax-M1, the world’s first open-source large-scale inference model built on a hybrid architecture. This groundbreaking development promises to significantly enhance efficiency and accessibility in complex productivity scenarios, rivaling the performance of leading proprietary models while offering exceptional cost-effectiveness.

The M1 model boasts an impressive 1 million token context window for input and an 80,000 token output capacity for inference. This capability, eight times larger than DeepSeek R1, is powered by a novel hybrid architecture featuring a lightning attention mechanism, dramatically improving computational efficiency when processing long contexts and conducting deep reasoning.

M1 represents a significant leap forward in open-source AI, stated a MiniMax spokesperson. Its hybrid architecture and lightning attention mechanism allow it to perform complex tasks with significantly less computational power. For example, when performing 80,000 token deep reasoning, M1 requires only about 30% of the computing power of DeepSeek R1.

This efficiency translates to substantial cost savings in both training and deployment. MiniMax reports that the reinforcement learning phase of M1’s training was completed using just 512 H800 GPUs over three weeks, with a total rental cost of $537,400.

Beyond its architectural innovations, MiniMax has also developed a new reinforcement learning algorithm called CISPO (Clipped Importance Sampling Policy Optimization). CISPO improves learning efficiency by clipping importance sampling weights, rather than traditional token updates. According to MiniMax, CISPO demonstrates twice the convergence performance of other reinforcement learning algorithms, including ByteDance’s DAPO, and significantly outperforms DeepSeek’s earlier GRPO algorithm.

MiniMax rigorously evaluated M1 across 17 industry-standard benchmark datasets. The results demonstrate a clear advantage in complex, productivity-oriented scenarios such as software engineering, long context understanding, and tool usage. Specifically, MiniMax-M1-40k and MiniMax-M1-80k achieved scores of 55.6% and 56.0% respectively on the SWE-bench verification benchmark. While slightly behind DeepSeek-R1-0528’s 57.6%, these scores significantly surpass those of other open-source models.

Furthermore, M1’s million-token context window enables superior performance in long context understanding tasks. The model surpasses all other open-source models and even outperforms proprietary models like OpenAI o3 and Claude 4 Opus, ranking second globally, only narrowly behind Google Gemini 2.5 Pro.

The release of MiniMax-M1 marks a significant step towards democratizing access to powerful AI technology. By open-sourcing the model, MiniMax aims to foster innovation and collaboration within the AI community, accelerating the development of advanced applications across various industries.

Conclusion:

MiniMax’s M1 model represents a significant breakthrough in open-source AI, combining innovative hybrid architecture, efficient reinforcement learning algorithms, and a massive context window to deliver exceptional performance and cost-effectiveness. Its open-source nature promises to accelerate the development and deployment of advanced AI applications across various industries. Further research and development in hybrid architectures and efficient training methods will likely lead to even more powerful and accessible AI models in the future.

References: